|
--- |
|
language: |
|
- en |
|
- da |
|
tags: |
|
- text-generation-inference |
|
- transformers |
|
- unsloth |
|
- mistral |
|
- trl |
|
base_model: Mabeck/Heidrun-Mistral-7B-base |
|
datasets: |
|
- Mabeck/danish-OpenHermes |
|
- kobprof/skolegpt-instruct |
|
license: mit |
|
--- |
|
|
|
|
|
<img src="https://huggingface.co/Mabeck/Heidrun-Mistral-7B-chat/resolve/main/heidrun.jpeg" alt="Heidrun Logo" width="400"> |
|
|
|
# Model description |
|
Heidrun-Mistral-7B-chat is a chat-model based on [Heidrun-Mistral-7B-base](https://huggingface.co/Mabeck/Heidrun-Mistral-7B-base), finetuned on [danish-OpenHermes](https://huggingface.co/datasets/Mabeck/danish-OpenHermes) and [skoleGPT](https://huggingface.co/datasets/kobprof/skolegpt-instruct) for a instruction/chat format. |
|
|
|
It is a new SOTA Danish open-source LLM and shows very strong performance in logic and reasoning tasks. |
|
|
|
Heidrun-7B ranks 1st among Danish open-sourced LLMs based on the [ScandEval](https://scandeval.com/mainland-scandinavian-nlg/) benchmark and shares 1st place on all Scandinavian languages. |
|
|
|
# Benchmarks |
|
|
|
The following benchmarks have been tested using [ScandEval](https://github.com/ScandEval/ScandEval). Rankings don't include merged models and GPT4 and GPT3.5 ranks 1 and 2: |
|
|
|
- **MMLU-da**: 35.66%+-0.85%/51.68+-0.63%, ranks 3rd |
|
- **DANSK**: 50.80% ± 2.33% / 34.04% ± 1.76%, ranks 3rd= |
|
- **Hellaswag-da**: 29.18% ± 0.99%/ 46.64% ± 0.76%, ranks 4th |
|
|
|
# Datasets |
|
This model is trained on Danish instruction datasets [danish-OpenHermes](Mabeck/danish-OpenHermes) and [skoleGPT](https://huggingface.co/datasets/kobprof/skolegpt-instruct), which have not been safeguarded or alligned. |
|
|
|
Most of the data has been machine-translated and may contain incorrect responses. |
|
|
|
# Samples |
|
|
|
This model uses the ChatML format. Using other formats may degrade the model's performance. |
|
|
|
Samples from the model using ChatML format: |
|
``` |
|
"<s> <im_start>system |
|
Du er en AI-hjælper. Du får en opgave fra en bruger hvor du kort og præcist skal forklare svaret. |
|
<im_end> |
|
<im_start>user |
|
Hvilke 3 populære retter findes der i dansk madlavning? |
|
<im_end> |
|
<im_start>assistant |
|
1. Smørrebrød: Smørrebrød er en populær dansk ret, der består af brød med forskellige toppings som røget laks, rejer, ost eller andre ingredienser. |
|
|
|
2. Frikadeller: Frikadeller er en traditionel dansk ret lavet af kødboller af kalvekød, svinekød eller lam, krydret med krydderurter og kogt i smør eller olie. |
|
|
|
3. Rød kål: Rød kål er en populær side ret i dansk madlavning. |
|
<im_end></s>" |
|
``` |
|
|
|
``` |
|
"<s> <im_start>system |
|
Du er en AI-hjælper. Du får en opgave fra en bruger hvor du kort og præcist skal forklare svaret. |
|
<im_end> |
|
<im_start>user |
|
Hvilke 3 områder skal fokuseres på for at effektivt tackle klimaforandringer? |
|
<im_end> |
|
<im_start>assistant |
|
1. Energiproduktion: Fokus på vedvarende energikilder som sol, vind og vandkraft for at reducere afhængigheden af fossile brændstoffer og minimere drivhusgasemissioner. |
|
|
|
2. Transport: Fremme bæredygtige transportformer som offentlig transport, cykling og elektriske køretøjer for at reducere luftforurening og afhængigheden af fossile brændstoffer. |
|
|
|
3. Skovbevarelse og affaldshåndtering: Fremme skovbevarelse og bæredygtig skovdrift for at fange kuldioxid og reducere drivhusgasemissioner, samt fremme genbrug og reducere affaldsproduktion for at mindske mængden af drivhusgasemissioner fra affald. |
|
<im_end></s>" |
|
``` |
|
|
|
# Uploaded model |
|
|
|
- **Developed by:** Mabeck |
|
- **Finetuned from model :** Mabeck/Heidrun-Mistral-7B-base |
|
|
|
This mistral model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library. |
|
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth) |
